| xf.li | bdd93d5 | 2023-05-12 07:10:14 -0700 | [diff] [blame^] | 1 | /* Copyright (C) 2012-2016 Free Software Foundation, Inc. |
| 2 | |
| 3 | This file is part of the GNU C Library. |
| 4 | |
| 5 | The GNU C Library is free software; you can redistribute it and/or |
| 6 | modify it under the terms of the GNU Lesser General Public |
| 7 | License as published by the Free Software Foundation; either |
| 8 | version 2.1 of the License, or (at your option) any later version. |
| 9 | |
| 10 | The GNU C Library is distributed in the hope that it will be useful, |
| 11 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 12 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 13 | Lesser General Public License for more details. |
| 14 | |
| 15 | You should have received a copy of the GNU Lesser General Public |
| 16 | License along with the GNU C Library. If not, see |
| 17 | <http://www.gnu.org/licenses/>. */ |
| 18 | |
| 19 | /* Assumptions: |
| 20 | * |
| 21 | * ARMv8-a, AArch64 |
| 22 | * Unaligned accesses |
| 23 | * |
| 24 | */ |
| 25 | |
| 26 | #define dstin x0 |
| 27 | #define src x1 |
| 28 | #define count x2 |
| 29 | #define tmp1 x3 |
| 30 | #define tmp1w w3 |
| 31 | #define tmp2 x4 |
| 32 | #define tmp2w w4 |
| 33 | #define tmp3 x5 |
| 34 | #define tmp3w w5 |
| 35 | #define dst x6 |
| 36 | |
| 37 | #define A_l x7 |
| 38 | #define A_h x8 |
| 39 | #define B_l x9 |
| 40 | #define B_h x10 |
| 41 | #define C_l x11 |
| 42 | #define C_h x12 |
| 43 | #define D_l x13 |
| 44 | #define D_h x14 |
| 45 | |
| 46 | #include <sysdep.h> |
| 47 | |
| 48 | ENTRY_ALIGN (memcpy, 6) |
| 49 | |
| 50 | mov dst, dstin |
| 51 | cmp count, #64 |
| 52 | b.ge L(cpy_not_short) |
| 53 | cmp count, #15 |
| 54 | b.le L(tail15tiny) |
| 55 | |
| 56 | /* Deal with small copies quickly by dropping straight into the |
| 57 | * exit block. */ |
| 58 | L(tail63): |
| 59 | /* Copy up to 48 bytes of data. At this point we only need the |
| 60 | * bottom 6 bits of count to be accurate. */ |
| 61 | ands tmp1, count, #0x30 |
| 62 | b.eq L(tail15) |
| 63 | add dst, dst, tmp1 |
| 64 | add src, src, tmp1 |
| 65 | cmp tmp1w, #0x20 |
| 66 | b.eq 1f |
| 67 | b.lt 2f |
| 68 | ldp A_l, A_h, [src, #-48] |
| 69 | stp A_l, A_h, [dst, #-48] |
| 70 | 1: |
| 71 | ldp A_l, A_h, [src, #-32] |
| 72 | stp A_l, A_h, [dst, #-32] |
| 73 | 2: |
| 74 | ldp A_l, A_h, [src, #-16] |
| 75 | stp A_l, A_h, [dst, #-16] |
| 76 | |
| 77 | L(tail15): |
| 78 | ands count, count, #15 |
| 79 | beq 1f |
| 80 | add src, src, count |
| 81 | ldp A_l, A_h, [src, #-16] |
| 82 | add dst, dst, count |
| 83 | stp A_l, A_h, [dst, #-16] |
| 84 | 1: |
| 85 | RET |
| 86 | |
| 87 | L(tail15tiny): |
| 88 | /* Copy up to 15 bytes of data. Does not assume additional data |
| 89 | being copied. */ |
| 90 | tbz count, #3, 1f |
| 91 | ldr tmp1, [src], #8 |
| 92 | str tmp1, [dst], #8 |
| 93 | 1: |
| 94 | tbz count, #2, 1f |
| 95 | ldr tmp1w, [src], #4 |
| 96 | str tmp1w, [dst], #4 |
| 97 | 1: |
| 98 | tbz count, #1, 1f |
| 99 | ldrh tmp1w, [src], #2 |
| 100 | strh tmp1w, [dst], #2 |
| 101 | 1: |
| 102 | tbz count, #0, 1f |
| 103 | ldrb tmp1w, [src] |
| 104 | strb tmp1w, [dst] |
| 105 | 1: |
| 106 | RET |
| 107 | |
| 108 | L(cpy_not_short): |
| 109 | /* We don't much care about the alignment of DST, but we want SRC |
| 110 | * to be 128-bit (16 byte) aligned so that we don't cross cache line |
| 111 | * boundaries on both loads and stores. */ |
| 112 | neg tmp2, src |
| 113 | ands tmp2, tmp2, #15 /* Bytes to reach alignment. */ |
| 114 | b.eq 2f |
| 115 | sub count, count, tmp2 |
| 116 | /* Copy more data than needed; it's faster than jumping |
| 117 | * around copying sub-Quadword quantities. We know that |
| 118 | * it can't overrun. */ |
| 119 | ldp A_l, A_h, [src] |
| 120 | add src, src, tmp2 |
| 121 | stp A_l, A_h, [dst] |
| 122 | add dst, dst, tmp2 |
| 123 | /* There may be less than 63 bytes to go now. */ |
| 124 | cmp count, #63 |
| 125 | b.le L(tail63) |
| 126 | 2: |
| 127 | subs count, count, #128 |
| 128 | b.ge L(cpy_body_large) |
| 129 | /* Less than 128 bytes to copy, so handle 64 here and then jump |
| 130 | * to the tail. */ |
| 131 | ldp A_l, A_h, [src] |
| 132 | ldp B_l, B_h, [src, #16] |
| 133 | ldp C_l, C_h, [src, #32] |
| 134 | ldp D_l, D_h, [src, #48] |
| 135 | stp A_l, A_h, [dst] |
| 136 | stp B_l, B_h, [dst, #16] |
| 137 | stp C_l, C_h, [dst, #32] |
| 138 | stp D_l, D_h, [dst, #48] |
| 139 | tst count, #0x3f |
| 140 | add src, src, #64 |
| 141 | add dst, dst, #64 |
| 142 | b.ne L(tail63) |
| 143 | RET |
| 144 | |
| 145 | /* Critical loop. Start at a new cache line boundary. Assuming |
| 146 | * 64 bytes per line this ensures the entire loop is in one line. */ |
| 147 | .p2align 6 |
| 148 | L(cpy_body_large): |
| 149 | /* There are at least 128 bytes to copy. */ |
| 150 | ldp A_l, A_h, [src, #0] |
| 151 | sub dst, dst, #16 /* Pre-bias. */ |
| 152 | ldp B_l, B_h, [src, #16] |
| 153 | ldp C_l, C_h, [src, #32] |
| 154 | ldp D_l, D_h, [src, #48]! /* src += 64 - Pre-bias. */ |
| 155 | 1: |
| 156 | stp A_l, A_h, [dst, #16] |
| 157 | ldp A_l, A_h, [src, #16] |
| 158 | stp B_l, B_h, [dst, #32] |
| 159 | ldp B_l, B_h, [src, #32] |
| 160 | stp C_l, C_h, [dst, #48] |
| 161 | ldp C_l, C_h, [src, #48] |
| 162 | stp D_l, D_h, [dst, #64]! |
| 163 | ldp D_l, D_h, [src, #64]! |
| 164 | subs count, count, #64 |
| 165 | b.ge 1b |
| 166 | stp A_l, A_h, [dst, #16] |
| 167 | stp B_l, B_h, [dst, #32] |
| 168 | stp C_l, C_h, [dst, #48] |
| 169 | stp D_l, D_h, [dst, #64] |
| 170 | add src, src, #16 |
| 171 | add dst, dst, #64 + 16 |
| 172 | tst count, #0x3f |
| 173 | b.ne L(tail63) |
| 174 | RET |
| 175 | END (memcpy) |
| 176 | libc_hidden_builtin_def (memcpy) |